File size: 5,397 Bytes
5958fad
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151
152
153
154
155
156
157
158
159
160
"""

Training script for surgical instrument classification

"""

import os
import pickle
import cv2
import pandas as pd
import numpy as np
from utils.utils import extract_features_from_image, fit_pca_lda_transformer, train_svm_model


def train_and_save_model(base_path, images_folder, gt_csv, save_dir, n_components=100):
    """

    Complete training pipeline that saves everything needed for submission

    

    Args:

        base_path: Base directory path

        images_folder: Folder name containing images

        gt_csv: Ground truth CSV filename

        save_dir: Directory to save model artifacts

        n_components: Number of PCA components

    """
    
    print("="*80)
    print("TRAINING SURGICAL INSTRUMENT CLASSIFIER")
    print("="*80)
    
    # Setup paths
    PATH_TO_GT = os.path.join(base_path, gt_csv)
    PATH_TO_IMAGES = os.path.join(base_path, images_folder)
    
    print(f"\nConfiguration:")
    print(f"  Ground Truth: {PATH_TO_GT}")
    print(f"  Images: {PATH_TO_IMAGES}")
    print(f"  PCA Components: {n_components}")
    print(f"  Save directory: {save_dir}")
    
    # Load ground truth
    df = pd.read_csv(PATH_TO_GT)
    print(f"\nLoaded {len(df)} training samples")
    print(f"\nLabel distribution:")
    print(df['category_id'].value_counts().sort_index())
    
    # Extract features
    print(f"\n{'='*80}")
    print("STEP 1: FEATURE EXTRACTION")
    print("="*80)
    
    features = []
    labels = []
    
    for i in range(len(df)):
        if i % 500 == 0:
            print(f"  Processing {i}/{len(df)}...")
        
        image_name = df.iloc[i]["file_name"]
        label = df.iloc[i]["category_id"]
        
        path_to_image = os.path.join(PATH_TO_IMAGES, image_name)
        
        try:
            image = cv2.imread(path_to_image)
            if image is None:
                print(f"  Warning: Could not read {image_name}, skipping...")
                continue
            
            # Extract features with enhanced configuration
            image_features = extract_features_from_image(image)
            
            features.append(image_features)
            labels.append(label)
            
        except Exception as e:
            print(f"  Error processing {image_name}: {e}")
            continue
    
    features_array = np.array(features)
    labels_array = np.array(labels)
    
    print(f"\nFeature extraction complete!")
    print(f"  Features shape: {features_array.shape}")
    print(f"  Labels shape: {labels_array.shape}")
    print(f"  Feature dimension: {features_array.shape[1]}")
    
    # Apply PCA+LDA
    print(f"\n{'='*80}")
    print("STEP 2: HYBRID DIMENSIONALITY REDUCTION (PCA → LDA)")
    print("="*80)
    
    combined_params, features_reduced = fit_pca_lda_transformer(
        features_array, 
        labels_array, 
        n_pca_components=N_COMPONENTS
    )
    
    print(f"\n  Final dimension: {features_reduced.shape[1]}")
    print(f"  Compression ratio: {features_array.shape[1] / features_reduced.shape[1]:.1f}x")
    
    # Train SVM
    print(f"\n{'='*80}")
    print("STEP 3: TRAINING SVM CLASSIFIER")
    print("="*80)
    
    train_results = train_svm_model(features_reduced, labels_array)
    
    svm_model = train_results['model']
    
    print(f"\nTraining complete!")
    print(f"  Support vectors: {len(svm_model.support_)}")
    
    # Save model artifacts
    print(f"\n{'='*80}")
    print("STEP 4: SAVING MODEL ARTIFACTS")
    print("="*80)
    
    os.makedirs(save_dir, exist_ok=True)
    
    # Save SVM model
    model_path = os.path.join(save_dir, "multiclass_model.pkl")
    with open(model_path, "wb") as f:
        pickle.dump(svm_model, f)
    print(f"  ✓ Saved SVM model: {model_path}")
    
    # Save PCA parameters
    # Save combined PCA+LDA parameters
    params_path = os.path.join(save_dir, "pca_lda_params.pkl")
    with open(params_path, "wb") as f:
        pickle.dump(combined_params, f)
    print(f"  ✓ Saved PCA+LDA params: {params_path}")
    
    print(f"\n{'='*80}")
    print("TRAINING COMPLETE!")
    print("="*80)
    print(f"\nFinal Results:")
    print(f"  Train Accuracy: {train_results['train_accuracy']:.4f}")
    print(f"  Test Accuracy:  {train_results['test_accuracy']:.4f}")
    print(f"  Test F1-score:  {train_results['test_f1']:.4f}")
    print(f"\nFiles saved to: {save_dir}")
    print(f"\nNext steps:")
    print(f"  1. Create a 'utils' folder in your HuggingFace repository")
    print(f"  2. Copy utils.py into the 'utils' folder")
    print(f"  3. Copy script.py, multiclass_model.pkl, and pca_params.pkl to the repository root")
    print(f"  4. Create an empty __init__.py file in the 'utils' folder")
    print(f"  5. Submit to competition!")


if __name__ == "__main__":
    
    BASE_PATH = "C:/Users/anna2/ISM/ANNA/phase1a2"
    IMAGES_FOLDER = "C:/Users/anna2/ISM/Images"
    GT_CSV = "C:/Users/anna2/ISM/Baselines/phase_1a/gt_for_classification_multiclass_from_filenames_0_index.csv"
    
    SAVE_DIR = "C:/Users/anna2/ISM/ANNA/phase1a2/submission"
    
    # Number of PCA components
    N_COMPONENTS = 250 #can be adjusted 
    
    # Train and save
    train_and_save_model(BASE_PATH, IMAGES_FOLDER, GT_CSV, SAVE_DIR, N_COMPONENTS)